Neural-DINF: A Neural Network based Framework for Measuring Document Influence

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6004–6009 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6004 Neural-DINF: A Neural Network based Framework for Measuring Document Influence Jie Tan, Changlin Yang, Ying Li, Siliang Tang∗, Chen Huang, Yueting Zhuang Zhejiang University {tanjie95, yclzju, ying.li}@zju.edu.cn, {siliang, beside.huang, yzhuang}@zju.edu.cn Abstract Measuring the scholarly impact of a document without citations is an important and challeng- ing problem. Existing approaches such as Doc- ument Influence Model (DIM) are based on dy- namic topic models, which only consider the word frequency change. In this paper, we use both frequency changes and word semantic shifts to measure document influence by devel- oping a neural network based framework. Our model has three steps. Firstly, we train word embeddings for different time periods. Subse- quently, we propose an unsupervised method to align vectors for different time periods. Fi- nally, we compute the influence value of docu- ments. Our experimental results show that our model outperforms DIM. 1 Introduction Identifying the most influential articles is of great importance in many areas of research. It is often the case that we are increasingly exposed to nu- merous papers published every day. Research on influence evaluation can be applied to measure the scholarly impact of universities and research facil- ities. Besides, it helps researchers to distinguish valuable research work from a large number of sci- entific papers. The common approach of assessing an article’s research impact is to count the num- ber of explicit references to it. However, citations are often not available. For example, collections including blog posts and government documents adopt ideas proposed in the documents without ex- plicit references (Stringer et al., 2008; Macroberts and Macroberts, 2010). To identify influential articles without citations, Gerrish and Blei (2010) and Gerow et al. (2018) proposed probabilistic methods, which are based on dynamic topic models (Blei and Lafferty, 2006). ∗Corresponding Author. They aimed to identify influential articles by exam- ining the word frequency change over time. In this paper, we aim to use both word frequency changes and word semantic shifts on measuring document influence without citations. For our pur- pose, we propose a neural network based method called Neural-DINF, which stands for a Neural Net- work based Framework for measuring Document Influence. Our idea is that words that have seman- tic shifts across time contribute significantly to the influence of a document. Recent studies show that words whose word embeddings across different time periods diverge significantly are suspected to have semantic shifts (Kim et al., 2014; Kulkarni et al., 2015; Hamilton et al., 2016). Neural-DINF first generates static word embeddings in each time period by using Word2Vec (Mikolov et al., 2013b,a) independently, then aligns embeddings to the same vector space with an unsupervised method, subsequently calculates differences of the embeddings of many words across time to identify words that experience semantic shifts, finally measures the influence of a document by counting these crucial words. In summary, this paper makes the following main …

Content truncated to 500 words